An intelligent shield machine pose prediction framework: HO-TCN-BiLSTM-multi-head attention mechanisms model

Abstract

Owing to the geological circumstances’ intricacy and operational scenarios, the pose of the shield machine is highly prone to alteration, and its dynamic variations are exceedingly challenging to acquire in time. Therefore, this paper proposes an intelligent pose prediction method that integrates a Hippopotamus Optimization (HO) Algorithm-optimized temporal convolutional network (TCN) and bidirectional long short-term memory (BiLSTM) enhanced by multi-head attention (HO-TCN-BiLSTM-multi-head attention) mechanisms to effectively address this challenge. First, Pearson correlation analysis selects tunneling parameters strongly related to pose. These are then denoised, decomposed, and reconstructed using Improved Complete Ensemble Empirical Mode Deplication with Adaptive Noise (ICEEMDAN), which preserves key variation characteristics while reducing data complexity. Second, the HO-TCN-BiLSTM-multi-head attention model is subsequently constructed, where TCN uses dilated convolution technology to improve the model’s feature-capturing capabilities, BiLSTM extracts bidirectional sequential information, and the multi-head attention mechanism highlights influential features. The key hyperparameters—including learning rate, BiLSTM neuron count, attention keys, and regularization coefficient—are optimized using the HO algorithm. Finally, evaluated on data from Beijing Subway Line 10, the model demonstrates high prediction accuracy, offering reliable guidance for real-time parameter adjustment and decision-making support for further realizing the intelligent autonomous control of the shield machine.